Tencent has introduced the Hunyuan Hy3 model, which features a Mixture of Experts (MoE) architecture with 295 billion parameters. This model has demonstrated strong capabilities in agent operations and climbed to eighth place on OpenRouter's list of most popular models, indicating a shift in the company's approach to creating AI models by altering its training pipeline.
After two years during which Tencent was considered lagging in the race for Chinese foundational models, the company has reasserted itself with the release of Hunyuan Hy3. Despite the absence of a major presentation event, the model reached eighth place in the ranking of most popular models on OpenRouter just three days after free access was provided. Furthermore, demand for the WorkBuddy product utilizing this model reached 50% queue levels, necessitating an emergency expansion of capacity, and Tencent's shares rose by 4.82% on the launch day.
Hy3 utilizes a Mixture of Experts architecture totaling 295 billion parameters, with 21 billion parameters activated during inference. Instead of aiming for maximum scores in benchmarks, developers focused on practical engineering implementation, inference cost efficiency, and agent task reliability. In the SWE-bench Verified test, Hy3 scored 78 points, trailing Claude Fable 5, GLM-5.2, and DeepSeek V4 Pro. However, on the more complex SWE-bench Pro test, it achieved 57.9 points, surpassing DeepSeek V4 Pro. Compared among open models, Hy3 outperformed both GLM-5.1 and DeepSeek V4 Pro on WildClawBench.
A more significant aspect is how Tencent achieved these results. In September 2025, the company hired Yao Shunyu, a former researcher from OpenAI, as Chief AI Scientist, marking a strategic pivot. Yao reorganized the Hunyuan R&D system, restored the pre-training and reinforcement learning pipelines, expanded and cleaned the training data. He also changed the development sequence, starting with product requirements for office tasks, coding, and knowledge management before defining the model architecture. This approach ensures that product feedback directly influences model training, closing the loop between model development and its real-world application.
The external assessment looks promising: daily token consumption for Hy3 increased 20-fold during the preview, and the selection of Hy3 by active users in WorkBuddy grew sixfold. In July's software model usage rankings on OpenRouter, Hy3 entered the top four alongside GLM-5.2, MiniMax M3, and the DeepSeek V4 series. Performance across various agent frameworks demonstrates controlled accuracy fluctuations within 4%, reduced hallucination frequency, and improved stability during multi-turn interactions.
Industry practitioners characterize Hy3 as a strong second-tier model that is achieving first-tier status among cost-effective open models. It is particularly competitive in frontend development, coding, search, and tool calling. While it does not yet lead in high-cognitive tasks, it has carved out a clear niche for large-scale, cost-sensitive enterprise deployments. DeepSeek V4 Pro and GLM-5.2 still outperform Hy3 in several complex tests, but the gap has significantly narrowed compared to previous Hunyuan iterations. The real test will begin when the free access ends. Nevertheless, Hy3 has already accomplished what no previous Hunyuan iteration could: it convinced external developers to actively choose Tencent's model over alternatives and prompted Tencent's internal products to assign it real workloads. Whether this leads to a sustainable competitive position depends on Tencent's ability to maintain a consistent pace of iteration and expand the model's capabilities while maintaining a price advantage.
Kuaishou announced the restructuring of its video generation unit, Kling AI, planning to spin it off into a separate company and conduct an independent funding round of up to $3 billion USD. A mandatory Initial Public Offering (IPO) date has been set for October 30, 2031. This information was disclosed on the evening of July 2, 2026, according to a company statement analyzed by Times Finance.
As part of this restructuring, the assets and operations of Kling AI will be consolidated into Beijing Keling Intelligent Technology Co., Ltd. The new structure involves independently raising up to $3 billion. Initial investors contributed $2.028 billion, and another 15 investors added approximately $766 million. The remaining deficit of $206 million is expected to be covered during the 60-day subscription period.
Thanks to this funding round, Kling AI's valuation reached approximately $18 billion post-investment, based on the assumption that $3 billion represents a 16.67% increase in registered capital. This figure is lower than market rumors of a target initial valuation of $20 billion. Industry analysts attribute this adjustment not to decreased performance, but to a broader market correction, as AI investments are shifting from narrative-based valuation to fundamentals-based valuation.
Kuaishou founder and CEO Chen Yixiao expressed confidence in Kling AI achieving over 100% year-over-year revenue growth this year. The lineup of investors draws attention because three major Chinese internet giants participated simultaneously: Tencent, Alibaba Cloud, and Baidu. Despite all having their own video generation models—Tencent's Hunyuan Video with the WorkRally platform, Alibaba's Happy Horse 1.1, and Baidu's MuseSteam—they decided to invest in a competitor.
Industry experts view this as a collective acknowledgment that video generation models have become an essential infrastructural layer for creating next-generation content.
Kling AI has shown rapid progress since its launch in June 2024. Currently, version 3.0 is the world's first to achieve native direct output in 4K format with precise motion control. The technology has been used in the production of TV series, including virtual scenes for CCTV's 'Peace Year' and footage for Hollywood's 'David's House'. In 2025, Kling AI generated approximately 1.1 billion yuan in revenue, and in the first quarter of 2026, revenue reached 650 million yuan, representing a year-on-year growth of over 300%. Annual recurring revenue approached $500 million as of March 2026, quadrupling the previous year's figure of $100 million. Seventy percent of the revenue comes from overseas markets, primarily North America, with the dual-engine model formed by corporate API calls and consumer subscriptions.
The spin-off structure is particularly significant because Kling AI's $18 billion valuation exceeds 60% of Kuaishou's parent company market capitalization, even though Kling AI accounted for less than 1% of Kuaishou's total revenue in 2025. Analysts note that the market applies fundamentally different valuation metrics for generative AI companies compared to traditional internet platforms, creating a strong incentive for separation.
